acceptodds
Under review as a conference paper at ICLR 2027

Align What Forms the Update: Bilinear Factor Alignment for Cross-Model Task-Vector Transfer

Abstract

Adapting large-scale pre-trained models to specialized tasks has been one of the prevalent paradigms in modern machine learning, with fine-tuning being one of the most widely studied approaches. However, when a new version of a pre-trained model becomes available, expertise acquired through fine-tuning cannot be directly reused because it is tied to the parameterization of the original model, requiring another costly fine-tuning. To address this inefficiency, recent work seeks to transfer expertise across models in the form of task vectors—defined as the parameter difference between a fine-tuned model and its base model. However, the connection between task-vector formation and cross-model alignment remains underexplored. In this work, we revisit task-vector formation and derive an alignment principle directly from the structure of the parameter update. Specifically, each layer-wise parameter gradient factorizes as an outer product between an input-side activation and an output-side gradient, revealing the bilinear structure underlying the update. This structure exposes two distinct factor spaces and motivates aligning their coordinate systems across models. Building on our observation, we propose BIFA (Bilinear Factor Alignment), a training-free framework that separately aligns the activation and gradient factor spaces across source and target models to transfer task vectors. Across extensive vision and language-model benchmarks, BIFA consistently outperforms existing transfer methods across models that differ in width, depth, and pre-training configuration, further extending to cross-family LLM transfer with greater architectural mismatch.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Knowledge transfer, Transfer learning, Representation alignment

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fFZuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this work the authors propose BiFA, an approach for transferring task vectors across pretrained models with different widths, depths and pretrainings. The authors observe that layer-wise parameter updates have a bilinear structure involving input activations and output gradients, thus they propose to align these two factor spaces across source and target models. The source task vector is then transformed through input-side and output-side mappings estimated from a calibration set. The method is evaluated on vision and language benchmarks.

Critical strengths
  • The bilinear interpretation provides a clear and intuitive motivation for separately aligning activation and gradient spaces.
  • The manuscript is clear and well written and the method is relatively simple despite addressing a difficult setting.
  • The empirical evaluation is extensive, covering different width and depth changes and different architectures, with strong results.
Critical weaknesses
  1. Relationship to Theseus. The task vector transport follows a two-sided linear transformation, which is largely inspired by Theseus. The main methodological distinction lies in how the two correspondence maps are constructed, particularly the use of task-conditioned output gradients. The resulting formulation is meaningful and empirical effective, but the degree of methodological novelty over Theseus is limited.

  2. Depth/Layer matching is not sufficently clear. For different-depth models the paper states that each target layer is associated with a source layer through relative-depth matching and that the corresponding source task vector is transferred. However it is unclear how this is implemented in practice and how does it behave when several target layers corresnpond to the same source layer (e.g., ViT-L to ViT-B). In particular, does the method directly use the activations/gradients and task vector of the selected source layer to construct each target-layer, or is an intermediate depth-matched source task vector/model explicitly materialized? This may effect also the comparison with Theseus, since this competitor materializes the expanded source model.

Minor:

  • The approach requires labeled calibration examples so the meaning of "training-free"should be stated more precisely.
  • The motivation for the use of CCA instead of CKA should be better clarified since one is used as a mapping estimator and the other appears in the analysis of layer correspondence.
Questions for authors
  1. Could the authors clarify the exact implementation of relative-depth matching? When source and target depths differ, do they directly reuse the matched source layer's activations, gradients, and task vector to construct each target-layer update, or do they first materialize a depth-matched task vector/model? If this differs from the strategy used by Theseus, it would be valuable to compare BiFA and Theseus under both depth-transfer strategies, to disentangle the contribution of factor alignment from the depth-matching procedure.
4 comments

Dear Reviewer fFZu, We sincerely thank you for your careful review and constructive suggestions. We particularly appreciate your positive assessment of the clear and intuitive bilinear interpretation, the simplicity and clarity of our framework, and the extensive empirical evaluation across different architectures and model configurations. Your comments on the distinction from THESEUS and the depth-matching procedure have helped us further clarify the contributions and implementation of BiFA. We address your concerns and questions in detail below.


[W1] Relationship to Theseus

Answer: We thank the reviewer for the careful comparison. While BiFA and THESEUS share a two-sided linear transport structure, BiFA's central contribution lies in identifying alignment spaces from the structure of task-vector formation, rather than simply using task-conditioned output gradients.

We further examine the structural basis for this choice and provide additional subspace analysis in our response to Reviewer kdux (W1, Theoretical justification of signal choices). As revealed by the bilinear decomposition in Eq. 5, the task vector's column and row spaces are directly related to the spans of output gradients and input activations, respectively. Importantly, since alignment maps are constrained by the subspaces spanned by their calibration signals, effective transfer requires these signals to adequately capture task-vector-relevant directions. Our additional subspace analysis shows that initialization gradient subspaces capture substantially more task-vector energy than output-activation subspaces at matched ranks (33.98% vs. 15.79% at 10% rank).

Consistent with this analysis, Table 5 shows that activation–gradient alignment achieves the highest transfer performance across all three mapping estimators: Procrustes, Ridge, and CCA. Table 6 further supports the importance of task-conditioned gradients. Furthermore, our additional experiments across different source optimizers, larger-scale LLMs, and task-vector addition and negation demonstrate consistent improvements over THESEUS (Appendices B.1 and G.4–G.5). These consistent improvements across diverse settings further support our central motivation that identifying alignment spaces from the bilinear structure of task-vector formation enables effective and generalizable cross-model task-vector transfer.

[W2, Q1] Clarification and Evaluation of Depth-Matching Strategies

Answer: We thank the reviewer for pointing out this ambiguity. BiFA does not construct an intermediate depth-matched model. Instead, each target layer jj is assigned a source layer using i(j)=round(j(LA−1)/(LB−1))i(j)=\mathrm{round}(j(L_A-1)/(L_B-1)), as described in Section 5.2. The corresponding source task vector is reused, while the activation- and gradient-based alignment maps are estimated independently for each target layer. Consequently, multiple target layers may share the same source task vector but receive different aligned updates.

To further examine whether the improvements stem from the alignment method or the depth-matching procedure, we conduct a controlled comparison by applying both interpolation-based and direct layer matching to THESEUS and BiFA-P.

MethodDepth MatchingAccuracy (%)
THESEUSInterpolation (as in THESEUS)77.28
BiFA-PInterpolation (as in THESEUS)79.97
THESEUSRelative depth (as in Ours)76.25
BiFA-PRelative depth (as in Ours)84.03

BiFA-P consistently surpasses THESEUS under either matching strategy, with gains of 2.69 and 7.78 percentage points under interpolation and direct matching, respectively. This confirms that the advantage of bilinear factor alignment is not merely a consequence of the depth-matching choice.

Furthermore, the effect of depth matching differs between the two methods. While THESEUS shows only a modest difference between interpolation and relative depth matching (77.28% vs. 76.25%), relative depth matching improves BiFA-P by 4.06 percentage points. We hypothesize that this advantage stems from the interaction between relative depth matching and bilinear factor alignment. Specifically, directly aligning the original source activations and gradients with each target layer may yield more effective target-specific alignment maps than using statistics from interpolated intermediate layers.

We have clarified the layer-matching implementation in Section 5.2 (L405–407) and added the controlled comparison to Appendix G.3 (L1726–1752) of the revised manuscript.

[W3] Clarification of Training-Free Setting

Answer: We thank the reviewer for raising this point. We clarify that “training-free” refers to the absence of parameter updates or iterative optimization during task-vector transfer, rather than the absence of labeled data or gradient computation. BiFA uses a small labeled calibration set to perform a single forward–backward pass at the pre-trained initialization, extracting activations and gradients to estimate the alignment maps. We have clarified the definition in the revised manuscript (L70-71).

[W4] Clarification on CKA vs. CCA

Answer: We thank the reviewer for highlighting this point. CKA [G11] and CCA [G12] serve distinct purposes in our study. CKA summarizes representational similarity into a single scalar score, allowing us to assess whether different layers exhibit similar representation structures. However, this score does not specify how individual feature directions correspond across the two spaces, and therefore cannot directly define an alignment map. In contrast, CCA explicitly identifies pairs of canonical directions that maximize correlation between the two representation spaces. These paired directions provide coordinate correspondences from which a linear alignment map can be constructed. Accordingly, we use CKA to analyze layer-wise representational similarity, while employing CCA to estimate explicit cross-model mappings. We have clarified the distinct roles of CKA and CCA in Appendix E (L1349-1367) of the revised manuscript.

Thank you to the authors for clarifying the raised concerns. I will keep my positive score.

Regarding the performance of BiFA on different depth models, looks like that most of the gain is given by the depth matching strategy. Similar gain are not observed on Theseus. Do the authors have some qunatitative insights on the improvement of BiFA, for example the actual alignment values?

We thank the reviewer for the positive assessment and insightful follow-up.

To quantitatively investigate why depth-matching strategies affect BiFA and THESEUS differently, we evaluate functional alignment under both strategies for ViT-B/16 → ViT-L/14. Specifically, we measure the cosine similarity between feature changes induced by transferred task vectors and those induced by target-model fine-tuning, both relative to the same pretrained target model. This metric evaluates how closely the transferred updates reproduce the target's task-specific functional effect, consistent with our objective in Section 3.1.

MethodDepth MatchingFeature Cosine ↑Accuracy (%)
THESEUSInterpolation (as in THESEUS)0.71177.28
BiFA-PInterpolation (as in THESEUS)0.72079.97
THESEUSRelative depth (as in Ours)0.69876.25
BiFA-PRelative depth (as in Ours)0.73584.03

BiFA-P achieves higher functional alignment under both depth-matching strategies, consistent with its higher transfer accuracy. Furthermore, relative-depth matching improves both functional alignment and accuracy for BiFA-P, whereas both decrease for THESEUS. This suggests that the effectiveness of depth matching depends on the underlying transfer method.

To further investigate why BiFA-P benefits from relative-depth matching, we extend the output-gradient correspondence analysis from Appendix D.2 (Fig. C) to the depth-mismatched setting. Specifically, we measure the top-10% subspace overlap between the aligned source and target output gradients under both depth-matching strategies.

Depth Matching (BiFA-P)Gradient Subspace Overlap ↑
Interpolation (as in THESEUS)0.496
Relative depth (as in Ours)0.577

Relative-depth matching yields higher output-gradient subspace overlap (+16.3%), indicating stronger correspondence between the dominant gradient directions of the matched source and target layers. Together with the functional alignment results, this suggests that BiFA-P benefits particularly from relative-depth matching through improved gradient-space correspondence.

Importantly, BiFA-P already outperforms THESEUS under the same interpolation strategy, indicating that its advantage is not solely attributable to depth matching. We have added these analyses to Appendix G.3 (L1753-1763).

Thank you for the insightful question and discussions that have further solidified the quality of our work.

Reviewer_kduxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BiFA, a training-free method for transferring task vectors across pre-trained models that differ in width, depth, pre-training data, or model family. Noting that a layer’s gradient factorizes as an outer product of the input activation and the output-side gradient, the authors estimate two linear maps per layer: one from paired activations and one from paired task-conditioned gradients. Both come from a single forward–backward pass over a small calibration set at initialization. Experiments on CLIP ViTs, T5, and cross-family LLMs show consistent gains over the THESEUS baseline, with the orthogonal Procrustes variant performing best.

Critical strengths
  1. The method needs only a single forward–backward pass on a small calibration set and closed-form map estimation. The Procrustes variant needs no hyperparameter tuning (fixed α=1), yet it outperforms baselines tuned on validation data.

  2. BiFA outperforms THESEUS across width, depth, and pre-training mismatches, in both transfer directions, and across vision, NLP, and cross-family LLM settings, often by large margins (e.g., +18.7 vs. +7.5 for width scaling at N=100).

  3. Table 5 shows the activation–gradient configuration is best under all three estimators. This indicates the gains come from the choice of alignment signals rather than from a particular estimator. Fig. 2 and Table 6 further support the design choices of gradient timing and task conditioning.

Critical weaknesses
  1. The form Rout⊤τRinR_{\mathrm{out}}^{\top}\tau R_{\mathrm{in}} in Eq. 6 follows from linearity alone, regardless of whether τ\tau is a sum of outer products. Therefore, the claim that the bilinear decomposition "derives" the alignment rule is overstated. The substantive contribution is the choice of aligning the output side with gradients, which is supported only empirically by Table 5, without theoretical justification. Moreover, the factor structure of input-activation and output-gradient spaces has long been exploited in the second-order optimization literature, such as K-FAC [1]. Mapping weights between models of different sizes via two-sided linear transformations also has precedents in LiGO [2] and OT-based model fusion [3, 4]. The relationship to these works is not discussed.

  2. If xB=xARinx_B = x_A R_{\mathrm{in}} and the outputs correspond via RoR_o, the function-preserving transport is τB=Ro⊤τA(Rin+)⊤\tau_B = R_o^{\top}\tau_A (R_{\mathrm{in}}^{+})^{\top}. Furthermore, since output gradients transform contravariantly, a map estimated from gradients approximately corresponds to (Ro⊤)+(R_o^{\top})^{+} rather than RoR_o. Eq. 6 coincides with this transport only when the maps are (semi-)orthogonal. Thus, the ridge and CCA variants may be functionally inconsistent even under perfect alignment. The large advantage of Procrustes (76.40 vs. 69.69/66.79) may be better explained by this issue than by the spectral-preservation argument in App. D.3. In addition, RR⊤=IRR^{\top}=I cannot hold for wide-to-narrow transfer (dA>dBd_A > d_B), and how this case is handled is not explained.

  3. One of the key value of task vectors lies in composing or removing tasks through arithmetic operations such as addition and negation [5]. However, all experiments in this paper transfer a single task vector and measure accuracy only on the corresponding task. There are no experiments that add multiple transferred vectors in the target to build a multi-task model, nor experiments that subtract a transferred vector to remove a task.

References

[1] J. Martens and R. Grosse. Optimizing Neural Networks with Kronecker-factored Approximate Curvature. ICML 2015.

[2] P. Wang, R. Panda, L. Torroba Hennigen, P. Greengard, L. Karlinsky, R. Feris, D. D. Cox, Z. Wang, and Y. Kim. Learning to Grow Pretrained Models for Efficient Transformer Training. ICLR 2023.

[3] S. P. Singh and M. Jaggi. Model Fusion via Optimal Transport. NeurIPS 2020.

[4] M. Imfeld, J. Graldi, M. Giordano, T. Hofmann, S. Anagnostidis, and S. P. Singh. Transformer Fusion with Optimal Transport. ICLR 2024.

[5] G. Ilharco, M. T. Ribeiro, M. Wortsman, S. Gururangan, L. Schmidt, H. Hajishirzi, and A. Farhadi. Editing Models with Task Arithmetic. ICLR 2023.

Questions for authors

Q1. It would be helpful to see experiments in which task vectors from multiple tasks are each transferred to the target and then added to build a multi-task model (addition), and in which a transferred vector is subtracted to remove the corresponding task (negation). Since RoutR_{\mathrm{out}} in BiFA is estimated from gradients conditioned on task labels, a different map is used for each task. Therefore, it is not obvious whether the transferred vectors remain composable in the target. In these experiments, please report the performance on control tasks (e.g., ImageNet) as well as on the target tasks. As comparisons, in addition to the existing baselines, I recommend including the following works that, like BiFA, use a data budget: Trans-LoRA [6] and PTA [7].

Q2. For the ridge and CCA variants, do the results change if the transfer is performed in a form that is consistent from a function-preserving perspective, e.g., using (Rin+)⊤(R_{\mathrm{in}}^{+})^{\top} instead of RinR_{\mathrm{in}} on the input side? I am curious whether the gap to Procrustes narrows. In addition, RR⊤=IRR^{\top}=I cannot hold for wide-to-narrow transfer (dA>dBd_A > d_B). Please explain how the Procrustes constraint was applied in this case.

Q3. When a single source layer is matched to multiple target layers, as in ViT-B/16 →\rightarrow ViT-L/14, the same update appears to be applied multiple times. How was this handled?

Q4. In Fig. 2, I wonder whether the performance drop at later fine-tuning steps is because "the early phase dominates the task vector," or because the gradient signal weakens and becomes noisier as the loss decreases.

References

[6] R. Wang, S. Ghosh, D. Cox, D. Antognini, A. Oliva, R. Feris, and L. Karlinsky. Trans-LoRA: Towards Data-Free Transferable Parameter Efficient Finetuning. NeurIPS 2024.

[7] J. Park and W. Cho. Platonic Task Arithmetic. NeurIPS 2026.

9 comments
AuthorsOpenReview ↗

Response to Reviewer kdux [1/4]

Dear Reviewer kdux, We sincerely thank you for your insightful comments and detailed technical feedback. We particularly appreciate your recognition of BiFA's training-free efficiency, strong performance across diverse architectures and transfer settings, and the effectiveness of activation–gradient alignment across multiple estimators. Your suggestions regarding theoretical justification, functional consistency, and task-vector composability have helped us substantially strengthen our analysis and experimental validation. We provide detailed responses to your concerns and questions below.


[W1-1] Clarification on the role of bilinear factorization

Answer: We thank the reviewer for this important clarification. We also agree that the bilinear decomposition motivates the choice of factor spaces to align, rather than deriving the transport rule itself. Specifically, it identifies input activations and output gradients as the relevant factors, while functional consistency provides the basis for the revised transport formulation, as further detailed in our response to [W2-1]. We have revised the manuscript accordingly.

[W1-2] Theoretical justification of signal choices

Answer: We thank the reviewer for this important observation. Our choice of alignment signals is motivated by the bilinear structure of task-vector formation and the identifiability of alignment maps. From Eq. 5, the column and row spaces of the source task vector satisfy

$\operatorname{col}(\tau_A)\subseteq\mathcal G_A =\operatorname{span}{g_{A,t}^{\top}},\qquad \operatorname{row}(\tau_A)\subseteq\mathcal X_A =\operatorname{span}{x_{A,t}}.$

Under two-sided task-vector transport, the output-side transformation acts on col⁡(τA)\operatorname{col}(\tau_A), which lies within the span of output gradients. Similarly, the input-side transformation acts on row⁡(τA)\operatorname{row}(\tau_A), which lies within the span of input activations. This matters because a correspondence map estimated from paired calibration signals is determined only along the directions spanned by those signals. Therefore, effective task-vector transfer requires alignment signals that adequately capture the task-vector-relevant subspaces.

To examine whether the initialization signals used in BiFA capture these task-vector-relevant directions, we measure the source task-vector energy captured by the leading activation and gradient subspaces on both the input and output sides, using 100 calibration samples:

Ein(Z,r)=∥τAVZ,r∥F2/∥τA∥F2,Eout(Z,r)=∥VZ,r⊤τA∥F2/∥τA∥F2.E_{\mathrm{in}}(Z,r)=\lVert\tau_A V_{Z,r}\rVert_F^2/\lVert\tau_A\rVert_F^2,\qquad E_{\mathrm{out}}(Z,r)=\lVert V_{Z,r}^{\top}\tau_A\rVert_F^2/\lVert\tau_A\rVert_F^2.

where VZ,rV_{Z,r} contains the leading right singular vectors of initialization signal ZZ at rank ratio rr.

As shown in Fig. 2 (new) of the revision, output gradients consistently capture more task-vector energy than output activations across rank ratios of 5–30%, with 33.98% vs. 15.79% at 10% rank. Similarly, input activations capture more row-space energy than input gradients (47.12% vs. 20.67%), supporting the complementary choice of input-side alignment signals.

These findings are consistent with Table 5, where Grad–Act achieves the lowest transfer accuracy, while Act–Grad achieves the highest. Together, the subspace analysis and empirical results provide a principled justification for our choice of alignment signals. We have added this analysis to Section. 4.3 (L303-320) and Fig. 2 of the revised manuscript.

[W1-3] Additional Related Work

Answer: We thank the reviewer for highlighting these relevant connections. We agree that the bilinear gradient structure and two-sided linear transformations are not new in themselves. However, these works differ from BiFA in their objectives and how these structures are utilized.

K-FAC [G4] similarly exploits activations and output gradients, but uses their statistics to approximate curvature for second-order optimization. In contrast, BiFA uses these factors as cross-model alignment signals.

LiGO [G5] similarly employs two-sided linear transformations, but learns them to transform pretrained weights for model growth. In contrast, BiFA estimates transformations from activation–gradient correspondences to transfer task updates.

OT-based model fusion [G6, G7] similarly addresses parameter misalignment, but estimates neuron correspondences from weight or activation similarities to fuse models. In contrast, BiFA aligns activation and task-conditioned gradient spaces to transfer task updates.

Thus, our contribution is identifying activation–gradient factor correspondences as principled signals for training-free cross-model task-vector transfer. We validate this choice across CCA, Ridge, and Procrustes, rather than relying on a specific mapping method as in previous task transfer works. We clarified this in Section 2 (L108–113) and Section 4.1 (L205–209).

AuthorsOpenReview ↗

Response to Reviewer kdux [2/4]

[W2-1, Q2-1] Evaluation of Function-Preserving Task-Vector Transport

Answer: We thank the reviewer for highlighting this important distinction. The original transport rule follows from the bilinear factorization, applying the estimated activation and gradient correspondences to the source task vector: τ^B=Rout⊤τARin.\hat\tau_B=R_{\mathrm{out}}^\top\tau_A R_{\mathrm{in}}.

However, this factor-wise transformation does not generally guarantee functional preservation. As the reviewer notes, preserving the source update's functional effect requires a pseudoinverse-based transport rule for general non-orthogonal maps: τ^B=Rout+τA(Rin+)⊤.\hat\tau_B=R_{\mathrm{out}}^+\tau_A(R_{\mathrm{in}}^+)^\top.The two rules coincide for (semi-)orthogonal Procrustes maps (R+=R⊤R^+=R^\top), but generally differ for non-orthogonal estimators such as Ridge and CCA.

To investigate this, we re-evaluated Ridge and CCA using function-preserving transport with reverse-fitted correspondences. Since direct pseudoinversion can amplify noise in poorly conditioned directions, we employ truncated pseudoinverses for numerical stability. The results are summarized below.

MethodNOriginal TransportFunction-Preserving Transport
BiFA-C (CCA)10066.7967.60
BiFA-R (Ridge)10069.6976.02
BiFA-P (Procrustes)10076.4076.40

Ridge substantially improves its transfer performance under the revised formulation, suggesting that the function-preserving transport rule is better suited to its non-orthogonal correspondence maps. In contrast, CCA shows only a modest improvement. We hypothesize that CCA's remaining performance gap stems from its low-rank alignment. As detailed in Appendix C.1 (L1063–1073) of the revised manuscript, CCA constructs its mapping using only the top-kk canonical components with the highest cross-model correlations, which may limit its coverage of task-vector-relevant directions and consequently restrict the transferable components of the source task vector.

Importantly, as shown in Table 5 (Section 5.4) of the revised manuscript, aligning input activations and output gradients consistently outperforms activation-only alignment across all alignment estimators, even under function-preserving transport. These results confirm that our central finding—the importance of selecting the appropriate factor spaces for alignment—remains valid independently of the transport rule. Together with the improved performance of Ridge, these results further demonstrate the flexibility of BiFA across different correspondence estimators.

While the original transport rule follows from the factor-wise construction, the reviewer's function-preserving formulation more directly aligns with our objective of transferring the functional effect of task-specific updates across models. We have therefore adopted function-preserving transport as the general formulation, particularly for non-orthogonal alignment maps, while the two formulations coincide for (semi-)orthogonal Procrustes maps. We denote the function-preserving variants as BiFA-C† and BiFA-R†, respectively, and have updated the corresponding experimental results in the revised manuscript.

This formulation has been incorporated into Section 4.2 (L245–255), with its full derivation and a mathematical discussion of why the two transport rules differ provided in Appendices B.2 and B.3 (L984–1062), respectively. We have also revised the analysis previously presented in Appendix D.3 and incorporated a discussion of CCA's relatively lower transfer performance in Appendix C.1 (L1095–1097) of the revised manuscript.

[W2-2, Q2-2] Clarification of Procrustes for Wide-to-Narrow Transfer

Answer: We thank the reviewer for pointing out this missing clarification. In Appendix C.3, Eq. (S) assumes dA≤dBd_A \leq d_B, where the Procrustes mapping satisfies RR⊤=IdARR^\top=I_{d_A}. For dA>dBd_A>d_B, this constraint cannot hold due to the dimensionality mismatch. We therefore fit the alignment in the reverse direction, from the lower-dimensional target space to the higher-dimensional source space:

Q∗=arg⁡min⁡Q∥ZBQ−ZA∥F2,s.t.QQ⊤=IdB.Q^\ast=\arg\min_Q\lVert Z_BQ-Z_A\rVert_F^2,\quad \mathrm{s.t.}\quad QQ^\top=I_{d_B}.

Given the compact SVD ZA⊤ZB=UΣV⊤Z_A^\top Z_B=U\Sigma V^\top, the solution is Q∗=VU⊤Q^*=VU^\top. We then transpose this map to obtain the source-to-target correspondence:

$R_P=(Q^*)^\top=UV^\top, \qquad R_P^\top R_P=I_{d_B}.$

Thus, although the alignment is fitted in the reverse direction, task-vector transfer remains source-to-target. Unlike the width-expansion case, this mapping projects the source representation onto a lower-dimensional subspace without preserving all source directions.

We use this formulation for the wide-to-narrow transfer experiments reported in Table I, demonstrating its effectiveness in this setting. We have clarified this procedure in Appendix C.3 (L1147–1162) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer kdux [3/4]

[W3, Q1] Composability of Transferred Task Vectors

Answer:We thank the reviewer for this constructive suggestion, which helped us further strengthen our work. We conduct additional task addition and negation experiments to evaluate the composability of transferred task vectors.

Baseline Selection. Following the reviewer's suggestion, we include PTA [G8], which supports training-free transfer on the same CLIP backbones and tasks, using 100 unlabeled probe images per task. In contrast, Trans-LoRA [G9] relies on LLM-generated synthetic data for adapter transfer, making it unsuitable for direct comparison in the vision setting. We therefore include Trans-LoRA as an additional baseline in our cross-family LLM transfer experiments.

1. Task Addition

We evaluate multi-task merging on eight vision tasks, transferring task vectors from ViT-B/16 to the wider ViT-B/16-plus architecture using 100 calibration samples per task (N=100). We independently transfer eight task vectors using task-specific alignment maps and combine them using Task Arithmetic (TA) [G2], which linearly combines task vectors, or TIES-Merging (TIES) [G3], which mitigates parameter interference by resolving sign conflicts and removing redundant updates. We evaluate the merged models on the eight target tasks and ImageNet as a control task.

MethodTarget Avg. (Δ↑)ImageNet (Δ↓)
Zero-shot57.6669.16
Fine-tuned + TA82.55 (+24.89)65.94 (−3.22)
Fine-tuned + TIES83.78 (+26.12)65.83 (−3.33)
THESEUS + TA60.82 (+3.16)68.71 (−0.45)
THESEUS + TIES61.09 (+3.43)68.58 (−0.58)
PTA68.30 (+10.64)59.70 (−9.46)
BiFA-P + TA70.54 (+12.88)68.96 (−0.20)
BiFA-P + TIES72.88 (+15.22)68.75 (−0.41)

BiFA-P achieves the highest multi-task accuracy among transfer-based methods while exhibiting minimal forgetting on ImageNet, demonstrating that task vectors independently transferred using their corresponding task-conditioned alignment maps remain effectively composable with limited interference. In contrast, PTA and directly fine-tuned task-vector combinations exhibit substantially greater degradation on the ImageNet control task, highlighting BiFA-P's favorable balance between multi-task performance and knowledge preservation.

2. Task Negation

We further subtract transferred task vectors from the pre-trained target model to examine whether the corresponding task knowledge can be selectively suppressed. Using N = 100 calibration samples per task, we report the average accuracy across the eight target tasks and ImageNet.

MethodTarget Avg. (Δ↓)ImageNet (Δ↓)
Zero-shot57.6669.16
Fine-tuned21.50 (−36.16)65.32 (−3.84)
THESEUS52.05 (−5.61)68.82 (−0.34)
PTA35.01 (−22.65)67.85 (−1.31)
BiFA-P33.81 (−23.85)68.64 (−0.52)

BiFA-P achieves effective task suppression while largely preserving unrelated knowledge. Compared with THESEUS, BiFA-P provides substantially stronger task suppression, while compared with PTA, it achieves slightly greater suppression with less ImageNet degradation. Although direct fine-tuned task-vector negation produces the largest target-task reduction, it also causes considerably greater control-task degradation.

  1. Trans-LoRA baseline

We additionally compare BiFA-P with Trans-LoRA on the original cross-family LLM transfer setting (Llama-3.2-3B → Gemma-2-2B).

MethodNInstruction followingMathMulti lingualCodeAvg.
zero-shot-13.6827.9744.1216.1025.47
THESEUS20014.0419.7842.5816.4023.20
Trans-LoRA20024.5821.9940.2814.2125.27
BiFA_P20022.7329.5744.6518.2928.81

At N = 200, BiFA-P achieves the highest average performance, outperforming Trans-LoRA by 3.54 percentage points. While Trans-LoRA performs better on instruction following, BiFA-P achieves stronger results on Math, Multilingual, and Code tasks, demonstrating broader improvements across the evaluated domains.

We have added the task addition and negation experiments to Appendix G.5 (L1776–1827) and Trans-LoRA as an additional baseline in Table 2 of the revised manuscript.

Reviewer_UoT3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BIFA, a method for transferring fine-tuning task vectors across models with different widths, depths, and pre-training initializations. Starting from the observation that a linear layer's gradient is the outer product of an output-gradient vector and an input-activation vector, BIFA estimates separate source-to-target alignment maps for these two spaces using labeled calibration data and transports the source task vector via R_out^T tau_A R_in. Among the tested estimators (CCA, ridge, and Procrustes), the Procrustes variant (BIFA-P) performs best and avoids hyperparameter tuning on a validation set. Experiments on CLIP backbones show substantial gains over activation-only transport baselines such as THESEUS (e.g., 76.40% vs. 65.15% at N = 100 in Table 1), alongside smaller-scale language model experiments on T5-3B to T5-Large and Llama-3.2-3B to Gemma-2-2B.

Critical strengths
  1. The ablations cleanly isolate the benefit of aligning output gradients in addition to input activations. Holding the Procrustes estimator fixed, switching from activation-activation to activation-gradient alignment improves accuracy from 65.15% to 76.40% (Table 5), with consistent gains for CCA and ridge as well. The vision results hold across width, depth, and pre-training mismatches (Tables 1, 3, and 4), and Appendix G adds useful checks on reverse transfer, class composition, and calibration seed variance.

  2. The method makes practical use of an existing source update while only requiring local forward/backward statistics from the frozen models. BIFA-P is particularly appealing because it does not require validation-set hyperparameter search or scaling tuning. The diagnostics in the appendix (such as comparing initialization gradients vs. later-trajectory gradients and relative-depth matching) also help clarify why the alignment signal works in the tested vision tasks.

Critical weaknesses
  1. The comparison against direct target fine-tuning under the same data and compute budget is not convincing enough. Although BIFA is framed as "training-free," it still requires labeled calibration samples and backward passes on the target model. A natural question is what happens if we simply use those exact same labeled examples and the same total wall-clock budget to directly fine-tune the target model. Appendix G.2 (Table D) includes a one-step fine-tuning baseline on the same data, which BIFA-P easily beats (e.g., 78.53% vs. 67.16% at K = 5). However, a single gradient step is a very weak baseline. In Table I, extracting calibration statistics takes 24.28s, whereas BIFA-P's full preparation takes 78.41s once alignment and SVD map-solving costs are included. Matching only a single forward/backward pass ignores this map-solving overhead. To justify the practical value of cross-model transfer over direct adaptation, the paper needs a comparison against properly tuned few-step full fine-tuning and parameter-efficient fine-tuning (e.g., LoRA) given the same labeled data and total wall-clock time, along with peak memory comparisons. In addition, BIFA-C and BIFA-R use extra validation data for hyperparameter selection, which should be explicitly accounted for when comparing data budgets.

  2. The empirical evidence for generality is limited in both LLM scale and optimizer coverage. First, on the LLM side, while Section 5.2 tests Llama-3.2-3B to Gemma-2-2B and Appendix G.2 tests T5-3B to T5-Large with encoder linear probing, generative cross-family transfer is only shown on a single relatively small pair (3B to 2B) in one direction. There are no experiments at the 7B+ scale, nor are there LLM calibration seed variances or preparation time/memory breakdowns. It is hard to tell from a single 3B-to-2B setup (+3.34 points at N = 200) whether the approach scales well to larger LLMs. Second, the theoretical motivation in Eq. 5 assumes vanilla gradient descent, where the accumulated task vector is a direct sum of raw gradient outer products. Appendix B discusses momentum and AdamW, and includes a nice rank-one reconstruction check for AdamW (showing 0.972 cosine similarity and similar transfer accuracy). However, that decomposition uses optimizer-rescaled factors, whereas BIFA estimates its alignment maps from unscaled initialization activations and gradients. It remains unclear why maps fitted on raw initialization factors should reliably transport updates preconditioned by AdamW, and there is no controlled experiment comparing task vectors trained with different source optimizers (e.g., SGD vs. momentum vs. AdamW) under the same setup.

Questions for authors
  1. Given the exact same labeled calibration set and a matched end-to-end wall-clock budget (including statistics extraction and map computation), how does BIFA-P compare against well-tuned few-step full fine-tuning or LoRA on the target model? Please also report peak memory and clarify how validation data used by BIFA-C/R is counted.

  2. Does BIFA remain effective on larger LLMs (e.g., 7B scale or above) and across other model pairs or transfer directions? If scaling up is computationally prohibitive, please report preparation time, memory cost, and seed variance on the current LLM setups and tone down the broader generality claims.

  3. How sensitive is BIFA to the optimizer used to produce the source task vector? Holding the architectures, task, and calibration setup fixed, does transfer quality change when the source model is fine-tuned with vanilla SGD, SGD with momentum, AdamW, or Muon?

6 comments
AuthorsOpenReview ↗

Response to Reviewer UoT3 [1/3]

Dear Reviewer UoT3, We sincerely thank you for your thoughtful and constructive feedback. We particularly appreciate your recognition of our well-designed ablation studies, strong empirical results across diverse transfer settings, and the practical efficiency of our training-free framework. Your comments on computational efficiency, scalability, and optimizer robustness have helped us further strengthen our evaluation through additional experiments and analyses. We address your concerns and questions in detail in the responses below.


[W1-1, Q1-1] Matched-Budget Comparison with Fine-Tuning

Answer: We thank the reviewer for this suggestion. We compare BiFA-P against target-model fine-tuning and LoRA for both vision and language models, using the same labeled calibration samples and end-to-end wall-clock budgets matched to BiFA-P's runtime, including statistics extraction and alignment map computation.

  1. Vision Experiments

We conduct the vision experiments comparing both ViT-B/16 → ViT-B/16 and ViT-B/16 → ViT-L/14 transfer settings on a single NVIDIA RTX 3090 GPU.

MethodNWall-clock time (s)Peak memory (GB)StepsAVG Acc.
Fine-tuned (B/16)10078.41s6.21G16978.58
LoRA-rank 64 (B/16)10078.41s4.53G27577.74
BiFA_P (B/16)10078.41s5.73G-79.66
Fine-tuned (L/14)100177s16.71G10783.50
LoRA-rank 64 (L/14)100177s11.58G11181.31
BiFA_P (L/14)100177s10.61G-84.03

Under identical data and time budgets, BiFA-P achieves the highest accuracy while offering a favorable memory–accuracy Pareto trade-off in the ViT-B/16 target setting. When scaling to the larger ViT-L/14 target, BiFA-P achieves both the highest accuracy and the lowest peak memory among the compared methods.

  1. LLM Experiments

To examine whether these advantages extend to larger models, we additionally conduct comparisons on 2B and 9B LLMs, using one NVIDIA H200 GPU for the 2B model and two NVIDIA H200 GPUs for the 9B model. In the 9B setting, fine-tuning is performed with Fully Sharded Data Parallel (FSDP) due to the model's large memory requirements relative to the available GPU memory. Peak memory is measured as the maximum concurrent sum of GPU memory usage across the two GPUs.

MethodNWall-clock time (s)Peak memory (GB)StepsAvg Acc.
Fine-tuned (2b)200307s72.2G1728.7
LoRA-rank 64 (2b)200307s32.3G2326.8
LoRA-rank 1024 (2b)200307s54.8G1927.8
BiFA_P (2b)200307s59.8G-28.8
Fine-tuned (9b)200846s275.0G2045.2
LoRA-rank 64 (9b)200846s83.4G2343.4
LoRA-rank 2048 (9b)200846s199.1G1444.9
BiFA_P (9b)200846s193.8G-48.5

Rank-64 LoRA offers lower memory usage but reduced accuracy, motivating comparisons with high-rank LoRA at comparable memory budgets. BiFA-P achieves higher accuracy than high-rank LoRA in both 2B (+1.0 pp) and 9B (+3.6 pp) settings. Compared with full fine-tuning, BiFA-P achieves comparable accuracy with lower peak memory at 2B and improves both metrics at 9B. Together with the vision results, these findings establish BiFA-P as a favorable accuracy–memory trade-off option across vision and language models. We have added these matched-budget comparisons to Appendix H.2 (L1879–1925) of the revised manuscript.

[W1-2, Q1-2] Clarification of Validation Data Usage

Answer: We thank the reviewer for raising this point. In the original evaluation, BiFA-C/R and THESEUS [G1] follow the Task Arithmetic [G2] validation protocol, using 10% of the dataset's predefined training split (capped at 5,000 samples) exclusively for hyperparameter selection.

To further reduce the reliance on additional data, we addionally evaluate the influence of hyperparameter tuning using only the same (N=100) calibration samples, without requiring additional labeled data.

MethodNValidation-based TuningCalibration-only Tuning
BiFA-C10066.7965.72
BiFA-R10069.6969.32

Calibration-only tuning achieves performance comparable to validation-based tuning, indicating that additional validation data is not essential for effective transfer in this setting. Accordingly, the revised Table 1 reports calibration-only results for BiFA-C/R, while retaining the original validation-tuned THESEUS baseline (65.15) for consistency with its established evaluation protocol. We have included these results and clarified the validation data usage in Appendix F.3 (L1497–1498) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer UoT3 [2/3]

[W2-1, Q2] Scalability to Larger LLMs and Seed Stability

Answer: We thank the reviewer for this suggestion. To further evaluate scalability, we extend BiFA-P to a larger LLM pair, Llama-3.1-8B → Gemma-2-9B, covering four downstream domains.

1. Transfer to Larger LLMs

MethodNInstruction followingMathMulti lingualCodeAvg.
zero-shot-14.870.354.632.943.1
THESEUS20019.267.655.033.243.7
BiFA_P20032.371.155.235.448.5

BiFA-P remains effective at the 8B–9B scale, outperforming THESEUS by 4.8 percentage points on average. Notably, its improvement over the zero-shot baseline is larger in the new larger-scale setting (+5.4 vs. +3.3 pp). These findings provide further evidence of BiFA-P's generality across different LLM scales and cross-family model pairs.

2. Stability Across Random Seeds

We further evaluate BiFA-P across five random seeds on the original cross-family LLM transfer setup (Llama-3.2-3B → Gemma-2-2B), covering four downstream domains.

MethodNInstruction followingMathMulti lingualCodeAvg.
zero-shot-13.6827.9744.1216.1025.47
THESEUS20014.0419.7842.5816.4023.20
BiFA_P20022.4 ± 1.728.66 ± 0.844.6 ± 0.319.8 ± 1.828.87 ± 0.6

BiFA-P achieves consistent performance across seeds, with a small standard deviation of 0.6 percentage points in average accuracy. This indicates that the observed improvements are stable across calibration seeds, providing additional support for the reliability of the results. We have added the results and corresponding discussions for the larger-scale LLM experiments to Appendix G.4 (L1764–1775) and the random-seed stability analysis to Appendix G.1 (L1556–1564) of the revised manuscript.

[W2-2] Alignment under AdamW Preconditioning

Answer: We thank the reviewer for this insightful observation. As derived in Appendix B, the AdamW task vector can be approximately expressed using optimizer-rescaled activation and gradient factors:

$\tau_A \approx -\sum_t \eta_t (g_{A,t}\operatorname{Diag}(r_t))^\top (x_{A,t}\operatorname{Diag}(c_t)),$

where rtct⊤r_tc_t^\top is the rank-one approximation of the AdamW pre-conditioner Pt=1/(v^t+ϵ).P_t=1/(\sqrt{\hat v_t}+\epsilon). Appendix B validates this decomposition, achieving a cosine similarity of 0.972 with the full AdamW task vector.

To examine whether optimizer-induced rescaling affects the estimated alignment, we re-estimate the maps using rescaled source signals. Since BiFA estimates alignment maps only once through a single forward–backward pass at initialization, rather than replaying the optimization trajectory, a single representative preconditioner is needed to account for the step-dependent rescaling introduced by AdamW. We therefore use the final AdamW preconditioner, where v^T\hat v_T aggregates gradient statistics throughout fine-tuning via an exponential moving average (EMA). Following Appendix B, we extract its dominant rank-one component rc⊤rc^\top to obtain the corresponding activation and gradient scaling factors.

Specifically, we estimate:

$\begin{aligned} \widetilde R_{\mathrm{in}}^{(l)} &=\mathcal A\left( \bar X_{A,\mathrm{pre}}^{(l)}D_c^{(l)}, \bar X_{B,\mathrm{pre}}^{(l)} \right),
\widetilde R_{\mathrm{out}}^{(l)} &=\mathcal A\left( \bar G_{A,\mathrm{pre}}^{(l)}D_r^{(l)}, \bar G_{B,\mathrm{pre}}^{(l)} \right). \end{aligned}$

We rescale only the source signals because the purpose is to account for the optimizer-induced scaling that formed the source task vector. The target signals remain unscaled, as the transferred update is applied directly in the target's original parameter coordinates without target-side task-specific optimization.

Using the same full AdamW task vector τA\tau_A, the rescaled alignment achieves 76.44%, compared with 76.40% using raw alignment. This suggests that BiFA's alignment is robust to optimizer-induced rescaling, supporting the use of raw activation and gradient signals for estimating cross-model correspondences. We have added this analysis to Appendix B (L909–951) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer UoT3 [3/3]

[W2-3, Q3] Robustness to Source Optimizers

Answer: We thank the reviewer for this insightful question. We evaluate BiFA-P using source task vectors obtained with four different optimizers on the eight vision tasks, using ViT-B/16 as the source model and ViT-B/16-plus as the target model, while keeping the architectures, tasks, and calibration setup fixed (N = 100 calibration samples).

MethodVanilla SGDSGD with momentumAdam WMuon
Target fine-tuned91.2892.7392.9092.87
THESEUS63.3864.8065.1563.24
BiFA_P72.7074.8976.4078.11

First, BiFA-P consistently outperforms THESEUS across all four optimizers, suggesting that alignment maps estimated from raw activations and gradients at pretrained initialization capture cross-model correspondences that remain effective for transferring task vectors produced by different source optimizers. Notably, BiFA-P achieves further improvements under Muon, unlike THESEUS, narrowing the performance gap to the oracle fine-tuned model.

Second, to better understand this improvement under Muon, we examine the spectral conditioning of the source task vectors by analyzing their singular-value distributions:

OptimizerEffective Rank ↑Top-10 Energy ↓
SGD2872.5%
Momentum SGD5659.0%
AdamW14039.6%
Muon5075.9%

Effective rank measures the number of effectively contributing singular directions (maximum 768), while top-10 energy measures the fraction of total energy captured by the ten largest singular values.

The analysis above suggests that vanilla SGD produces a highly concentrated spectrum, while momentum SGD and AdamW yield progressively flatter spectra. Muon exhibits the most distributed spectrum, consistent with its orthogonalized updates. As discussed in Appendix C.3 (L1109–1119), the semi-orthogonal Procrustes mappings preserve the geometry of the source representation without introducing anisotropic scaling, thereby preserving the non-zero singular values of the source task vector under width-expanding transport. Accordingly, concentrated task vectors may be more sensitive to alignment errors in a few dominant directions, whereas more distributed updates may be less affected by errors in individual directions as suggested by strong performance of BiFA-P with Muon optimizer.

Together with our AdamW preconditioning analysis (76.40% vs. 76.44% using raw and rescaled alignment), these results suggest that BiFA-P consistently outperforms THESEUS regardless of the source optimizer, while transfer quality is closely associated with the spectral conditioning of the source task vector. We have added the optimizer robustness results and the corresponding spectral analysis to Appendix B (952–982) in the revision. We appreciate the reviewer's suggestion, which helped us strengthen the empirical evaluation and deepen our understanding of BiFA's robustness across different source optimizers.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank the reviewers (UoT3, kdux , fFZu) for their time, careful consideration, and constructive feedback. Their comments have helped us clarify key aspects of our work and strengthen the paper. This global response contains a summary of the strengths highlighted by the reviewers, revisions to the manuscript, and references cited in our responses. We address the weaknesses and questions raised by each reviewer separately in our individual responses. We look forward to further discussion during the discussion period.

Highlighted Strengths of our work

  1. [UoT3 , fFZu] Clear and intuitive presentation with a simple, well-motivated framework.
  2. [fFZu] Bilinear interpretation of task-vector formation provides intuitive motivation for factor-wise alignment.
  3. [UoT3 , kdux] Well-designed ablations validate activation–gradient alignment across three mapping estimators.
  4. [UoT3 , kdux ,fFZu] Practical with strong empirical results under extensive evaluation: diverse architectures, model sizes, and transfer settings.
  5. [UoT3, kdux] Training-free and efficient transfer using a single forward–backward pass and closed-form alignment.
  6. [UoT3, kdux] Particularly appealing with no hyperparameter tuning required for BiFA-P, while outperforming tuned baselines.
  7. [UoT3, kdux] Comprehensive analyses of gradient timing, task conditioning, and transfer robustness.

Major Revisions of the manuscript

For ease of reference, the changes made in the revised manuscript are highlighted in blue.

  1. Page 3–4, Lines 108–113, 205–209: We expanded the related work discussion to clarify the distinction between BiFA and prior approaches, including OT-based model fusion, LiGO, and K-FAC. [kdux]
  2. Page 5, Lines 245–255; Pages 19–20, Lines 984–1061: We adopted function-preserving transport as the general formulation, provided its full derivation in Appendix B.2, discussed its differences from direct factor-wise transport in Appendix B.3, and updated the affected Ridge and CCA results. [kdux]
  3. Pages 5–6, Lines 303–319: We clarified the motivation for initialization gradients and added a task-vector subspace analysis to justify the selection of activation and gradient alignment signals. [kdux]
  4. Pages 17–19, Lines 910–980: We added experiments on optimizer-induced rescaling and source optimizer robustness across SGD, momentum, AdamW, and Muon, together with spectral analysis. [UoT3]
  5. Pages 24–25, Lines 1273–1340: We added an effective sample size analysis to investigate the concentration of gradient contributions across fine-tuning steps and explain the later-step performance drop. [kdux]
  6. Page 8, Lines 405–408; Pages 32–33, Lines 1727–1752: We clarified the implementation of relative-depth matching and added experiments comparing depth-matching strategies for BiFA and THESEUS. [fFZu, kdux]
  7. Pages 29–30, Lines 1556–1573; Page 33, Lines 1764–1775: We added LLM calibration-seed robustness experiments and extended cross-family transfer to larger 8B–9B models. [UoT3]
  8. Pages 33–34, Lines 1776–1827: We added task addition and negation experiments to evaluate the composability of transferred task vectors, including comparisons with PTA. [kdux]
  9. Pages 34–35, Lines 1879–1925: We added matched-budget comparisons against target-model fine-tuning and LoRA, including end-to-end runtime and peak memory analysis. [UoT3]

Global Reference

[G1] Rinaldi et al., “Transporting Task Vectors across Different Architectures without Training,” ICML 2026.

[G2] Ilharco et al., “Editing Models with Task Arithmetic,” ICLR 2023.

[G3] Yadav et al., “TIES-Merging: Resolving Interference When Merging Models,” NeurIPS 2023.

[G4] Martens and Grosse, “Optimizing Neural Networks with Kronecker-factored Approximate Curvature,” ICML 2015.

[G5] Wang et al., “Learning to Grow Pretrained Models for Efficient Transformer Training,” ICLR 2023.

[G6] Singh and Jaggi, “Model Fusion via Optimal Transport,” NeurIPS 2020.

[G7] Imfeld et al., “Transformer Fusion with Optimal Transport,” ICLR 2024.

[G8] Park and Cho, “Platonic Task Arithmetic,” NeurIPS, 2026.

[G9] Wang et al., “Trans-LoRA: towards data-free Transferable Parameter Efficient Finetuning,” NeurIPS, 2024.

[G10] Zhou et al., “On Task Vectors and Gradients,” UniReps Workshop, PMLR 2026.

[G11] Kornblith et al., “Similarity of Neural Network Representations Revisited,” ICML 2019.

[G12] Hotelling, “Relations Between Two Sets of Variates,” Biometrika 1936.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.